SADE3: an effective system for automated detection of epileptiform events in long-term EEG based on context information

• Published in: Medical & biological engineering & computing 2006-06, Vol.44 (6), p.459-470
• Main Authors: Argoud, Fernanda I M, De Azevedo, Fernando M, Neto, José Marino, Grillo, Eugênio
• Format: Article
• Language: English
• Subjects: Automation; Databases, Factual; Electroencephalography; Electroencephalography - methods; Epilepsy; Epilepsy - diagnosis; Humans; Neural networks; Neural Networks (Computer); Pattern recognition; Signal Processing, Computer-Assisted; Wavelet transforms
• ISSN: 0140-0118; 1741-0444
• DOI: 10.1007/s11517-006-0056-y

Interictal spike detection is a time-consuming, low-efficiency task, but is important to epilepsy diagnosis. Automated systems reported to date usually have their practical efficacy compromised by elevated rates of false-positive detections per minute, which are caused mainly by the influence of artifacts (such as noise activity and ocular movements) and by the adoption of single or simple approaches. This work describes the development of a hybrid system for automatic detection of spikes in long-term electroencephalogram (EEG), named System for Automatic Detection of Epileptiform Events in EEG (SADE(3)), which uses wavelet transform, neural networks and artificial intelligence procedures to recognize epileptic and to reject non-epileptic activity. The system's pre-processing stage filters the EEG epochs with the Coiflet wavelet function, which showed the closest correlation to epileptogenic (EPG) activity, in opposition to some other wavelet functions that did not correlate with these events. In contrast to current attempts using continuous wavelet transform, we chose to work with fast wavelet transform to reduce processing time and data volume. Detail components at appropriate decomposition levels were used to accentuate spikes, sharp waves, high-frequency noise activity and ocular artifacts. These four detailed components were used to train four specialized neural networks, designed to detect and classify the EPG and non-EPG events. An expert module analyzes the networks' outputs, together with multichannel and context information and concludes the detection. The system was evaluated with 126,000 EEG epochs, obtained from seven different patients during long-term monitoring, under diverse behavior and mental states. More than 6,721 spikes and sharp waves were previously identified by three experienced human electroencephalographers. In these tests, the SADE(3) system simultaneously achieved 70.9% sensitivity, 99.9% specificity and a rate of 0.13 false-positives per minute, indicating its usefulness and low vulnerability to artifact influence. After tests, the SADE(3) system showed itself to be able to process bipolar cortical EEG records, from long-term monitoring, up to 32 channels, without any data preparation or event positioning. At the same time, SADE(3) revealed a high capacity to reject non-epileptic paroxysms, robustness in relation to a variety of spike morphologies, flexibility in adjustment of performance rates and the capacity to actually sa